Methodology
Know what is behind the data.
From a study question to a release your team can inspect. Each step has a defined output.
Start with the task.
Specify what the model should do, what evidence it may see and what would count as an acceptable result. The collection follows that specification.
Study brief
Population, sources, acquisition routes, geography, collection windows and participant burden.
Acceptance plan
Required fields, coverage, reference labels, review requirements and the treatment of missing or rejected observations.
Establish what can be collected.
Assess source access, legal and research requirements, recruitment and qualified review before a real pilot. A data type appearing in the catalogue does not establish that it is available.
Source and permission frameworkKeep the source and both clocks.
An event can occur before it becomes available to a model. Preserve both timestamps, alongside the source artifact, collection route and transformations.
Measured and reported
Distinguish device measurements, participant accounts, extracted values and reviewer judgments.
Missing and conflicting
Keep gaps explicit. Hold conflicting sources for review instead of inventing a complete timeline.
Corrected
Link a correction to the earlier entry and preserve the history of what was known when.
Make the review decision inspectable.
A review protocol should define admissible evidence, reviewer qualifications, calibration and the route for disagreement. Clinical review is a separately staffed task, not a label applied to every file.
Source checks
Confirm integrity, evidence pointers, units, timing and documented transformations.
Annotation checks
Record instructions, reviewer decisions, uncertainty and adjudication. Measure agreement only when the review has actually been run.
Acceptance
Report coverage, missingness and exclusions against the agreed criteria. Preserve rejected examples separately from accepted data.
Deliver the explanation with the data.
A usable release includes its schema, source manifest, quality report, label guide, split rules and permitted-use schedule. Participant separation and temporal access rules depend on the task.
Training and evaluation
Define population splits and contamination checks before making performance claims.
Outcome labels
Specify how a later observation is established. Follow-up alone does not show that an intervention caused it.
Versioned package
Track file integrity, release versions, corrections and access conditions.